What I Built In My First Week of Claude Code Club

I joined the club at $9 a month on a Sunday night and told myself the only rule for the week ahead was that I had to ship one thing a day. Not finish, not polish, not perfect - just ship. A live URL, a working extension, a script that ran. The whole point of the week was to feel the loop rather than read about it. Below is the honest account of what got built, what broke, and what shifted in my head along the way.

Quick context on where I started. I'd written maybe twenty lines of HTML in my life, had never opened a terminal except to follow a tutorial step by step, and could not have told you what a Git commit was without Googling it. I'd also been around enough AI tools to know the conversational ones felt useful but rarely produced anything I shipped. So when I opened Claude Code on day one, my expectations were tempered. By day seven they were not.

One thing I'll flag up front: every single project below started ugly. The day-one landing page had Comic-Sans-energy. The inbox cleaner deleted the wrong email on the first run. The Chrome extension froze my whole browser and made me force-quit. None of that is in the highlight reel anyone else posts. I'm leaving it in because the ugly middle is where the actual learning happened, and pretending it didn't exist would be the most useless version of this post I could write.

The first build was a one-page personal site at my own name. Bio, links, a contact button. I sat down at 8 p.m., asked Claude Code to create a single page with three sections and a clean look, and watched it scaffold an entire project structure I didn't understand. Twenty minutes in I had something running locally that looked like a real website. The font was wrong, the spacing was off, and the contact button did nothing - but it was a real page in a real browser, and that was the bar.

What I learned in those first two hours wasn't a technical skill, it was the cadence. Ask for one small thing, run it, look at it, ask for the next small thing. The temptation when something works is to immediately ask for five more features at once. The temptation when something breaks is to delete everything and start over. Both impulses kill the rhythm. The win on day one was deploying that ugly page to a free hosting service before I went to bed, so I had a real URL to send myself a screenshot of in the morning.

One small win on day one - I wrote a four-line CLAUDE.md at the root of the project that said what the page was, who the audience was, and what tone to use. I did not understand at the time how much that one file was doing in the background. Every later session, Claude already knew context I would otherwise have had to re-explain. That was probably the single best decision of the whole week, and I made it almost by accident.

Day two I got ambitious and wired up a Gmail MCP server to clean my inbox. The plan was simple - find every unread newsletter from the last 90 days and archive it. The first version of the script worked, in the sense that it ran. It also archived three personal emails I needed because my filter for "newsletter" was sloppier than I thought. I dug them out of the archive, calmed down, and wrote down the lesson: when you give Claude reach into a real account, the cost of being imprecise jumps from "a wasted hour" to "oops, I just acted on three months of mail."

I rewrote the script with two guards. First, a dry-run mode that printed every action it would take without actually taking it. Second, a five-email limit on each run so a bug couldn't cascade into a disaster. With those guards in place, the cleaner did its job - about 340 newsletters archived over the next hour, none of them anything I needed. The clean inbox felt great. The real win was internalizing that any automation touching a real account needs guardrails baked in from the first version, not added after the first scare.

The middle of the week I tried to build a Chrome extension. I'd seen people post about tiny extensions that did one useful thing, and I wanted one that highlighted certain keywords on any page I visited. Claude Code generated the manifest, the content script, the popup - everything looked right. I loaded it as an unpacked extension and within ten seconds my browser went unresponsive. The script was running an infinite loop on every page load. I force-quit, dug through the code with Claude, found the runaway loop, fixed it, and reloaded. Browser back. Lesson learned.

Day four was finishing it. The extension worked, but the UI was an eyesore - a default-styled popup with a single textarea. I spent the evening iterating on small visible slices: rounded the corners, added padding, changed the font, made the highlight color configurable. None of those changes were technically hard, but they each taught me something about the loop. The big realization was that I'd spent two days mostly fighting the same kind of small UI issues, not because they were complex, but because I kept asking for them in batches of three or four at a time. The moment I started asking for one tweak at a time and running between each, the progress accelerated.

By the end of day four I had a Chrome extension I actually used. It's not a product, it doesn't make money, and nobody else has installed it. But it's mine, it lives in my browser, and it works. That sounds like a tiny thing. After three days of believing I'd never be able to build a browser extension, it did not feel tiny at all.

Day five was when the picture started to compose. I'd been talking to a friend who runs a small consultancy and has to send weekly status reports to four clients. Same shape every week - what we shipped, what's blocking, what's next. He hated it. I asked if I could try to automate it and he said yes, more out of curiosity than confidence. Two evenings later he had a script that pulled the week's commits and notes from his project tool, drafted the report in his voice, and saved it to a doc he could review before sending.

The build itself was nothing fancy - read from one MCP server, prompt for the report, write to another. What made it click was watching him use it. The first draft was almost right. The second iteration, after he gave me three pieces of feedback, was indistinguishable from the reports he'd been hand-writing. He shipped it to his actual clients the next week. The moment I realized something I'd built was being used by a real person to do real work, the whole week reframed. It wasn't "I'm learning Claude Code." It was "I built a thing someone else uses." Those are different sentences.

I didn't charge him for it - it was a favor and I learned more from the build than he could have paid me for - but I noted what the equivalent project would cost on a real invoice. Based on what I've read inside the club, a productized version of that report bot is in the $1,500 to $2,400 range as a one-off, or about $400 a month as a retainer. I'm not running an agency yet. But day five was the first time I could honestly imagine running one.

Day seven was reflection, not building. I spent an hour going back through the week - opening each project, reading the CLAUDE.md files, looking at the shape of the prompts I'd sent. A few things stood out clearly enough that I want to flag them, because they're the lessons that would compress the next person's week if they took them seriously from day one.

The moment Claude Code stopped feeling like magic and started feeling like a tool was somewhere around day four, halfway through fixing the Chrome extension. The magic phase is fun but it doesn't survive contact with real problems - you can't be in awe of the same tool that just froze your browser. What replaces awe, if you stick with the loop, is something better: a calm, working trust that you can ask for something specific and probably get it. That's the shift. That's why I'm still here at $9 a month, and that's why I'd do the first week again exactly the same way.